Accurate Prostate Cancer Detection and Segmentation Using Non-Local Mask R-CNN With Histopathological Ground Truth.

Abstract

PURPOSE/OBJECTIVE(S)\nWe aim to develop deep learning (DL) models to accurately detect and segment intraprostatic lesions (IL) on biparametric MRI (bp-MRI).\n\n\nMATERIALS/METHODS\nThree patient cohorts with ground truth IL delineated on different modalities were collected. 158 patients from two datasets had suspicious ILs delineated based on bp-MRI: 97 patients were from PROSTATEx-2 Challenge with biopsy result independent from bp-MRI based delineation, 61 patients were from IMPROD clinical Trial with biopsy done for each delineation; 64 patients from IMPROD clinical Trial had ILs identified and delineated by using whole mount prostatectomy specimen sections as reference standard; 40 private patients were unlabeled. We proposed a non-local Mask R-CNN to improve segmentation accuracy by addressing the imperfect registration issue between MRI modalities. We also proposed to post aggregate 2D predictions to estimate IL volumes within the whole prostatic gland and keep top-5 3D predictions for each patient. In order to explore the small dataset problem, we employed different learning techniques including transfer learning and semi-supervised learning with pseudo labelling. We experimented with two label selection strategies to see how they affected model performance. The first strategy kept only one prediction by referring to biopsy result, in order to minimize false positives; while the second strategy kept all top-5 predictions. 3D top-5 detection rate, dice similarity coefficient (DSC), 95 percentile Hausdorff Distance (95 HD, mm) and true positive ratio (TPR) were our evaluation metrics. We compared DL model prediction with prostatectomy-based ground truth delineation to accurately evaluate the boundary and malignancy level. We separately evaluated ILs of all Gleason Grade Group (GGG) and clinically significant ILs (GGG > 2).\n\n\nRESULTS\nMain results are summarized in Table 1.\n\n\nCONCLUSION\nOur proposed method demonstrates state-of-art performance in the detection and segmentation of ILs and shows great effectiveness for clinically significant ILs.